Behavior of tropopause height and atmospheric temperature in models, reanalyses, and observations: Decadal changes

Transcription

1 JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 108, NO. D1, 4002, doi: /2002jd002258, 2003 Behavior of tropopause height and atmospheric temperature in models, reanalyses, and observations: Decadal changes B. D. Santer, 1 R. Sausen, 2 T. M. L. Wigley, 3 J. S. Boyle, 1 K. AchutaRao, 1 C. Doutriaux, 1 J. E. Hansen, 4 G. A. Meehl, 3 E. Roeckner, 5 R. Ruedy, 4 G. Schmidt, 4 and K. E. Taylor 1 Received 1 March 2002; revised 21 June 2002; accepted 24 June 2002; published 3 January [1] We examine changes in tropopause height, a variable that has hitherto been neglected in climate change detection and attribution studies. The pressure of the lapse rate tropopause, p LRT, is diagnosed from reanalyses and from integrations performed with coupled and uncoupled climate models. In the National Centers for Environmental Prediction (NCEP) reanalysis, global-mean p LRT decreases by 2.16 hpa/decade over , indicating an increase in the height of the tropopause. The shorter European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis has a global-mean p LRT trend of 1.13 hpa/decade over Simulated p LRT trends over the past several decades are consistent with reanalysis results. Superimposed on the overall increase in tropopause height in models and reanalyses are pronounced height decreases following the eruptions of El Chichón and Pinatubo. Interpreting these p LRT results requires knowledge of both T(z), the initial atmospheric temperature profile, and T(z), the change in this profile in response to external forcing. T(z) has a strong latitudinal dependence, as does T(z) for forcing by well-mixed greenhouse gases and stratospheric ozone depletion. These dependencies help explain why overall tropopause height increases in reanalyses and observations are amplified toward the poles. The pronounced increases in tropopause height in the climate change integrations considered here indicate that even AGCMs with coarse vertical resolution can resolve relatively small externally forced changes in tropopause height. The simulated decadal-scale changes in p LRT are primarily thermally driven and are an integrated measure of the anthropogenically forced warming of the troposphere and cooling of the stratosphere. Our algorithm for estimating p LRT (based on a thermal definition of tropopause height) is sufficiently sensitive to resolve these large-scale changes in atmospheric thermal structure. Our results indicate that the simulated increase in tropopause height over is a robust, zero-order response of the climate system to forcing by well-mixed greenhouse gases and stratospheric ozone depletion. At the global-mean level, we find agreement between the simulated decadal-scale p LRT changes and those estimated from reanalyses. While the agreement between simulated p LRT changes and those in NCEP is partly fortuitous (due to excessive stratospheric cooling in NCEP), it is also driven by real pattern similarities. Our work illustrates that changes in tropopause height may be a useful fingerprint of human effects on climate and are deserving of further attention. INDEX TERMS: 1610 Global Change: Atmosphere (0315, 0325); 1620 Global Change: Climate dynamics (3309); 3362 Meteorology and Atmospheric Dynamics: Stratosphere/troposphere interactions; 8409 Volcanology: Atmospheric effects (0370); KEYWORDS: tropopause height, tropospheric temperatures, stratospheric temperatures, climate modeling, microwave sounding unit Citation: Santer, B. D., et al., Behavior of tropopause height and atmospheric temperature in models, reanalyses, and observations: Decadal changes, J. Geophys. Res., 108(D1), 4002, doi: /2002jd002258, Program for Climate Model Diagnosis and Intercomparison, Lawrence Livermore National Laboratory, Livermore, California, USA. 2 Institut für Physik der Atmosphäre, Deutsches Zentrum für Luft- und Raumfahrt, Oberpfaffenhofen, Wessling, Germany. 3 National Center for Atmospheric Research, Boulder, Colorado, USA. 4 NASA Goddard Institute for Space Studies, New York, New York, USA. 5 Max-Planck Institute for Meteorology, Hamburg, Germany. Copyright 2003 by the American Geophysical Union /03/2002JD Introduction [2] In 1996, the Intergovernmental Panel on Climate Change (IPCC) concluded that [Houghton et al., 1996, p. 4] the balance of evidence suggests a discernible human influence on global climate. The latest IPCC report reinforces this conclusion, and finds that [Houghton et al., 2001, p. 10] most of the observed warming over the last 50 years is likely to have been due to the increase in ACL 1-1

2 ACL 1-2 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT greenhouse gas concentrations. These statements are based on multiple lines of evidence. One line of evidence is from formal statistical studies that seek to identify model-predicted fingerprints of anthropogenic effects in observational climate records [Santer et al., 1996a; Mitchell et al., 2001]. [3] To date, this work has focused on either surface temperature [Santer et al., 1995; Hegerl et al., 1996, 1997; Wigley et al., 1998; Tett et al., 1999; Stott et al., 2000; Allen et al., 2002] or on zonally averaged temperatures in the free atmosphere [Karoly et al., 1994; Santer et al., 1996b; Tett et al., 1996]. The case for a discernible human influence on global climate would be strengthened by successful fingerprint studies of variables other than surface temperature or zonal-mean temperature profiles. A global-scale human-induced climate change signal should be manifest in a variety of climate variables, and not in temperature alone. [4] Examples of statistical studies using alternative detection variables include analyses of changes in integrated ocean heat content [Barnett et al., 2001] and Northern Hemisphere sea ice extent [Vinnikov et al., 1999]. There has also been interest in various temperature-related indices, such as the temperature contrast between land and ocean, low and high latitudes, and the Northern and Southern Hemispheres [Karoly and Braganza, 2001]. Other indices, such as the Greenhouse Climate Response Index of Karl et al. [1996] and the Common-Sense Climate Index of Hansen et al. [1998], simultaneously consider changes in temperature and moisture. This body of work suggests that a coherent human-induced climate change signal is emerging in a number of different variables, and is not limited to temperature only. [5] One variable that has not been examined from a climate change standpoint is the height of the tropopause. The tropopause is the transition zone between the turbulently mixed troposphere and the more stably stratified stratosphere [Hoinka, 1998]. This transition is also marked by changes in lapse rate and the chemical composition of the atmosphere. Thus the location of the tropopause can be defined in a variety of ways, according to dynamic, thermal, and chemical properties of the atmosphere [Hoinka, 1998; Seidel et al., 2001]. [6] Various studies have attempted to elucidate the key factors that determine the latitude-altitude distribution of the tropopause [Held, 1982; Thuburn and Craig, 1997; Haynes et al., 2001]. Initial investigations were based on radiosonde data [Reid and Gage, 1984, 1985]. More recently, analyses from numerical weather prediction centers [Reichler et al., 1996] and reanalysis products [Hoinka, 1998; Randel et al., 2000] have provided insights into the climatological properties of the tropopause, its seasonal cycle, and its secular variations on interannual and decadal timescales. Much of this work has been summarized by Seidel et al. [2001]. It reveals that the tropopause responds to a variety of influences, such as variations in solar radiation [Reid and Gage, 1981], atmospheric angular momentum [Reid and Gage, 1984], stratospheric ozone [Dameris et al., 1995; Hoinka, 1998], and explosive volcanic eruptions [Reid and Gage, 1985; Randel et al., 2000]. [7] There have been few model-based studies of tropopause height. Thuburn and Craig [1997, 2000] used an atmospheric General Circulation Model (GCM) to test different theories regarding the key factors determining tropopause height. They found the simulated tropopause height to be sensitive to surface temperature (through changes in the atmospheric moisture distribution), somewhat less sensitive to the distribution of ozone and changes in the equator-to-pole temperature gradient, and relatively insensitive to changes in the Earth s rotation rate. Since these experiments were performed in perpetual January mode, possible controls on decadal-timescale changes in tropopause height could not be examined. [8] A number of observational studies have identified decadal-timescale changes in tropopause height. A useful review is given by Ramaswamy et al. [2001]. Virtually all of these studies show an increase in the height of the tropopause over the last several decades or longer, either at individual radiosonde stations [Steinbrecht et al., 1998], networks of stations in the tropics [Seidel et al., 2001], or in tropical averages from reanalysis products [Randel et al., 2000]. One obvious question is whether similar changes occur in climate model experiments driven by estimates of historical changes in natural and anthropogenic forcings. [9] Ramaswamy et al. [2001] suggest that the poor vertical resolution of most current climate models precludes their simulation of multi-decadal changes in tropopause height. To test this assumption, we investigate tropopause height changes in climate change experiments performed with two different coupled atmosphere-ocean GCMs [Roeckner et al., 1999; Hansen et al., 2002]. We find that despite relatively coarse vertical resolution, both models show similar decadal-scale increases in tropopause height in response to anthropogenic forcing. [10] To gain insights into the possible controls on lowfrequency changes in tropopause height, we examine its behavior in control integrations and in climate change experiments with different combinations of anthropogenic and natural forcings. The forcings that we focus on are changes in well-mixed greenhouse gases, stratospheric and tropospheric ozone, volcanic aerosols, and direct and indirect sulfate aerosol effects. Variations in these constituents (either individually or jointly) can yield markedly different temperature responses in the troposphere and stratosphere [Hansen et al., 1997, 2002; Bengtsson et al., 1999]. Such large-scale changes in the static stability of the atmosphere may in turn influence tropopause height. Thus by studying the stratospheric and tropospheric temperature responses to different forcing mechanisms, we can help to elucidate the effects of these forcings on tropopause height. The layeraveraged stratospheric and tropospheric temperatures that we compute here are also a key component of a companion paper (B. D. Santer et al., Behavior of tropopause height and atmospheric temperature in models, reanalyses, and observations: Signal detection, submitted to Journal of Geophysical Research, 2002) (hereinafter referred to as Santer et al., submitted manuscript, 2002) which seeks to compare the relative detectability of changes in temperature and tropopause height. [11] The structure of the present paper is as follows. In section 2, we introduce the model and observational data sets used here and describe the methods employed for calculating layer-averaged temperatures and tropopause pressure. Section 3 gives a brief comparison of the climatological mean

3 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT ACL 1-3 state of tropopause height in model and reanalysis data. The remainder of the paper focuses on decadal-scale changes in tropopause height and layer-averaged atmospheric temperature. Global-mean changes in these variables are compared in section 4. The comparison focuses on the last two decades, although there is also some analysis of simulated changes over the next years. In section 5, we examine how greenhouse-gas forcing modifies spatially averaged profiles of atmospheric temperature and affects tropopause height. Horizontal patterns of changes in atmospheric temperature and tropopause height are described in section 6. Simple pattern correlations are used to assess the relative influence of stratospheric and tropospheric temperature fluctuations on tropopause height. Section 7 provides a brief summary and conclusions. 2. Data and Calculation of Derived Variables 2.1. Observational Data [12] Our temperature data are from the satellite-based MSU instrument, which uses the upwelling microwave radiation from oxygen molecules to estimate the vertically weighted temperature of deep atmospheric layers [Spencer and Christy, 1990]. Most interest in MSU temperatures has focused on products for the stratosphere, mid- to upper troposphere, and lower troposphere. These are monitored by (respectively) MSU channels 4 and 2, and by the lower tropospheric retrieval (2LT). The MSU weighting functions for these three layers peak at 74, 595, and 740 hpa. [13] We rely on version d of the MSU data [Christy et al., 2000]. Data were available as monthly means for latitude/longitude gridboxes. They have full global coverage and span the 264-month period January 1979 through December [14] We obtain estimates of observed tropopause pressure and equivalent MSU temperatures from reanalysis projects, which involve use of an atmospheric numerical weather forecast model with a fixed observational data assimiliation system. We employ data from two reanalyses. The first was jointly performed by the National Centers for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR) [Kalnay et al., 1996]. The second was conducted by the European Centre for Medium-Range Weather Forecasts (ECMWF) [Gibson et al., 1997]. We refer to these reanalyses subsequently as NCEP and ERA. [15] Detailed descriptions of these data sets are given elsewhere [Pawson and Fiorino, 1998; Santer et al., 1999]. Here, it is sufficient to note that the operational models had different horizontal resolution (T62 for NCEP versus T106 for ERA), but similar vertical resolution (17 pressure levels). Sixteen of these 17 levels are identical (10, 30, 50, 70, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925 and 1000 hpa). NCEP has a pressure level at 20 hpa that is absent in ERA, while ERA has a 775 hpa level that is not present in NCEP. To facilitate comparison of the reanalyses, they were interpolated to a common latitude/longitude grid. [16] The NCEP data extend from January 1948 through December We did not use NCEP temperatures prior to January 1979 due to well-documented inhomogeneities that exist around the time of transition to satellite data assimilation [Pawson and Fiorino, 1998; Santer et al., 1999; Randel et al., 2000]. The ERA reanalysis does not span this transition period, and extends from January 1979 through February Even after 1979, however, it is likely that climate records from both reanalyses incorporate discontinuities related to biases in the assimilated satellite data [Trenberth et al., 2001] Model Data [17] We used model data from three sources: the Max- Planck Institute for Meteorology in Hamburg (MPI), NCAR, and the Goddard Institute for Space Studies (GISS). The models and integrations in question have been described in detail elsewhere [Roeckner et al., 1996, 1999; Bengtsson et al., 1999; Washington et al., 2000; Meehl et al., 2000; Hansen et al., 2002; Santer et al., 2001]. [18] The MPI data were from experiments performed with the ECHAM4 atmospheric GCM coupled to an isopycnal ocean model (OPYC; Oberhuber [1993]). The ECHAM4 model was jointly developed by MPI and the Meteorological Institute of the University of Hamburg on the basis of an ECMWF forecast model. The AGCM was run at T42 spectral truncation and has a 19-level hybrid sigma-pressure coordinate system, with the top of the model atmosphere at 10 hpa. Temperature data used in the calculation of tropopause height were archived at 15 discrete pressure levels (30, 50, 70, 100, 150, 200, 250, 300, 500, 600, 700, 850, 900, 950, and 1000 hpa). The OPYC ocean component has comparable horizontal resolution to the atmosphere poleward of 36, with gradually decreasing meridional grid spacing equatorward of 36, down to 0.5 at the equator. OPYC has 11 vertical layers. [19] In the present study, we analyze six different ECHAM4/OPYC integrations. The baseline integration, GHG, has time-dependent changes in well-mixed greenhouse gases. It commences in 1860 and ends in Other forcings are added sequentially to the baseline case. The additional forcings are the direct scattering effects of sulfate aerosols (GSD), indirect aerosol effects and tropospheric ozone changes (GSDIO), stratospheric ozone depletion (GSO1, GSO2), and stratospheric aerosols from the June 1991 eruption of Mount Pinatubo (GSOP). Full details of the forcings are given by Roeckner et al. [1999] and Bengtsson et al. [1999]. [20] GSD and GSDIO start in 1860 and extend to 2050, while GSOP, GSO1 and GSO2 span the 228-month period from January 1979 through December The GSO1 and GSO2 experiments use January 1979 oceanic initial conditions from GSDIO. GSO1 and GSO2 have the same forcing, but start from slightly different atmospheric initial conditions in January GSOP s initial conditions were taken from June 1991 in GSO1. We also analyzed ECHAM4/OPYC data from a 300-year control integration with no changes in natural or anthropogenic forcings. [21] In addition to the ECHAM control run, we also considered data from a 300-year control integration performed with the Parallel Climate Model (PCM) jointly developed by NCAR and the Los Alamos National Laboratory [Washington et al., 2000]. The atmospheric component of PCM is version 3 of NCAR s Community Climate Model (CCM3; Boville and Hurrell [1998]) with a T42 spectral truncation (roughly horizontal resolu-

4 ACL 1-4 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT tion) and 18 vertical levels. The PCM ocean component has 32 vertical levels and 2/3 2/3 horizontal resolution, decreasing to 0.5 at the equator. [22] Our use of PCM data provides a second model-based estimate of the natural internal variability of tropopause height and atmospheric temperature. We employ this information to assess whether the externally forced signals in these variables are large relative to their unforced variability (section 4.3). The PCM variability estimates are also a key component of the optimal detection scheme used by Santer et al. (submitted manuscript, 2002). [23] A key result of our study is that there are pronounced tropopause height signals in climate change experiments performed with the ECHAM model (section 4.3). To investigate the model dependence of this finding, we examine tropopause height changes simulated by the GISS SI2000 AGCM [Hansen et al., 2002]. Our analysis of GISS climate change simulations is limited to global-mean tropopause height data only. We stress that it serves primarily as an independent check on the tropopause height signals estimated from ECHAM. [24] The GISS SI2000 AGCM has coarser horizontal and vertical resolution (4 5 latitude/longitude and 12 vertical layers) than either ECHAM or PCM. The AGCM employs sigma coordinates up to 150 hpa, and pressure coordinates above 150 hpa, with discrete pressure level boundaries at 10, 30, 60, 100, 150, 210, 285, 390, 550, 720, 854, 934, and 984 hpa. We used a model version with 2nd order differencing in the momentum equation. In the experiments considered here, the SI2000 AGCM was driven by prescribed, timevarying changes in observed sea surface temperatures (SSTs) and sea ice, as well as by combined temporal variations in well-mixed greenhouse gases, volcanic aerosols, solar irradiance, ozone, stratospheric water vapor, and tropospheric aerosols [Hansen et al., 2002]. There are five different realizations of this experiment, each commencing from slightly different atmospheric initial conditions. Although the integrations cover the period , our analysis is restricted to only to facilitate comparisons with ECHAM. The integrations, forcings, and key physics improvements in the atmospheric model (relative to the GISS SI95 AGCM) are described in detail by Hansen et al. [2002] Calculation of Derived Variables [25] To facilitate comparison with the actual MSU temperatures, we used a static global-mean weighting function to compute equivalent MSU temperatures from the reanalysis, ECHAM, and PCM data. This procedure is described by Santer et al. [1999]. For global and hemispheric means, the static weighting function results are very similar to those obtained with a full radiative transfer code. [26] We use the thermal definition of the tropopause here, also referred to as the lapse rate tropopause (LRT) [Hoinka, 1998; Seidel et al., 2001]. The thermal tropopause has been defined by the World Meteorological Organization (WMO) as [World Meteorological Organization (WMO), 1957, p. 137] the lowest level at which the lapse rate decreases to 2 C/km or less, provided also the average lapse rate between this level and all higher levels within 2 km does not exceed 2 C/km. Relative to other definitions of the tropopause, such as those based on potential vorticity (PV), the thermal definition has the advantage that it is easily computed from sigma-level or pressure-level temperature data. The PV-defined tropopause height cannot be computed close to the equator and depends on the (somewhat arbitrary) choice of the threshold PV value [Hoinka, 1998]. The LRT has the disadvantage that it is sometimes difficult to compute at very high latitudes, for reasons discussed by Reichler et al. [1996]. [27] The algorithms used to derive LRT [cf. Reichler et al., 1996; Hoinka, 1998] are typically applied using linear interpolation of the lapse rate in a p k -coordinate system, where p denotes pressure, k = R/c p, and R and c p are the gas constant for dry air and the specific heat capacity of dry air at constant pressure. Differences between algorithms relate to the number of levels involved in the interpolation. Hoinka [1998] identifies the model level that first satisfies the WMO criterion, and then computes the mean lapse rates of the four layers above and below the two central layers surrounding the threshold level. He defines the pressure of the LRT (henceforth p LRT ) as the point at which the mean lapse rates intersect. In contrast, the interpolation by Reichler et al. [1996] relies only on the layers immediately above and below the threshold level. [28] We use the Reichler et al. [1996] algorithm here, and apply this to ECHAM, PCM, and reanalysis data sets. At each grid-point, we calculate p LRT from the monthly mean temperature profile at discrete pressure levels. Our justification for the use of monthly mean data to compute p LRT is discussed in Appendix A. Global-mean p LRT values for the GISS climate change experiments were also computed with the WMO criterion, using an algorithm very similar to that of Reichler et al. [1996]. 3. Climatological-Mean Tropopause Height [29] Our focus here is on studying decadal-scale changes in p LRT, and on identifying the key forcings that may be responsible for such changes. In the following, we provide only a brief comparison of the mean state of p LRT in model and reanalysis data. [30] Meridional profiles of zonally and annually averaged p LRT are shown in Figure 1 for the reanalyses, ECHAM, and PCM. All have a higher tropopause in the tropics and subtropics, a much lower tropopause toward the poles, and steep gradients in tropopause pressure between the subtropics and high latitudes [cf. Hoinka, 1998]. [31] The NCEP and ERA results are very similar except at mid- to high latitudes in the Southern Hemisphere, where radiosonde coverage is sparse and the reanalyses are poorly constrained by observations. PCM s tropopause is consistently lower than in both reanalyses between roughly 25 N 25 S. In contrast, ECHAM s tropopause is higher than in NCEP and ERA between approximately 20 N 20 S. In the vicinity of the subtropical jets, ECHAM s p LRT errors are larger than those in PCM, which may reflect a smaller cold bias in PCM. ECHAM overestimates tropopause height poleward of 40 in both hemispheres. PCM shows similar but less pronounced errors. [32] The differences between simulated zonal-mean p LRT values and those in NCEP and ERA generally exceed the differences between the reanalyses themselves. Model-ver-

5 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT ACL 1-5 Figure 1. Zonal-mean annual-mean p LRT in reanalyses and model integrations. The averaging periods are listed in the legend. p LRT values are not shown poleward of 70 due to difficulties in defining p LRT at high latitudes [Reichler et al., 1996]. The grey shading marks the ±2s limits for the ECHAM control run p LRT values. It is a measure of how p LRT varies due to (simulated) interannual variability. The interannual variability in p LRT for PCM (not shown) is similar to that of ECHAM. sus-reanalysis discrepancies cannot be explained by external forcing alone, since they are larger than the p LRT differences between the ECHAM control run and the GSOP experiment. They are also larger than the uncertainties in estimates of the simulated mean state arising from interannual variability in p LRT (see grey shading in Figure 1). [33] Note that p LRT was estimated from ECHAM, PCM, and reanalysis datasets in which the vertical resolution was similar (section 2). We infer from this that the differences between simulated climatological mean p LRT values and those in NCEP and ERA are not simply due to overall vertical resolution, and likely reflect both resolution differences in the vicinity of the tropopause and systematic model errors, such as cold biases in the vicinity of the polar night jets [Gates et al., 1999]. 4. Global-Mean Changes 4.1. Time Series [34] We begin by examining global-mean monthly mean anomaly time series of p LRT and layer-averaged atmospheric temperature. Our focus is on January 1979 through December 1997, the period covered by the three ECHAM experiments most relevant for comparison with observations and reanalyses (GSO1, GSO2, and GSOP). We also present some observational and reanalysis results for the longer period January 1979 through December Lapse Rate Tropopause [35] In the NCEP reanalysis, the most prominent feature of the global-mean p LRT data is a decrease in tropopause pressure of 2.16 hpa/decade over , corresponding to an increase in the height of the tropopause. Superimposed on this overall trend is higher-frequency variability, part of which is associated with the eruptions of El Chichón and Pinatubo (Figure 2a). Values of p LRT increase after both eruptions, indicating a lower tropopause altitude. The maximum increase after Pinatubo is larger than after El Chichón, consistent with the larger estimated forcing of Pinatubo [Hansen et al., 1997, 2002; Andronova et al., 1999; Santer et al., 2001; T. M. L. Wigley and B. D. Santer, Differential ENSO and volcanic effects on surface and tropospheric temperatures, submitted to Journal of Climate, 2002] (hereinafter referred to as Wigley and Santer, submitted manuscript, 2002). It is evident that the volcanically induced p LRT signals persist for at least several years. [36] Values of p LRT also show overall decreases in the ECHAM and GISS climate change experiments (Figures 2b and 2c). Consider the ECHAM results first. In GSO1, GSO2, and GSOP, the linear trends in p LRT over are 2.43, 1.94, and 1.47 hpa/decade (respectively). The p LRT trend in NCEP over the same time period is 1.82 hpa/decade. GSO1 and GSO2 have larger negative trends than NCEP due to their neglect of the p LRT increases caused by El Chichón and Pinatubo. GSOP has a smaller negative trend than NCEP due to its neglect of El Chichón and its slightly larger p LRT increase after Pinatubo. Inclusion of El Chichón s effects in GSOP would yield a larger negative trend in p LRT, bringing GSOP in better agreement with the NCEP result. At the global-mean level, the decay time of the p LRT response to Pinatubo is similar in GSOP and the reanalyses (Figure 2d). [37] In the five realizations of the GISS climate change experiment, p LRT trends over range from 1.30 to 1.52 hpa/decade (Figure 2c). The trend in the ensemble-mean of the GISS p LRT data is 1.40 hpa/decade. None of the GISS trend results is significantly different from the NCEP p LRT trend of 1.82 hpa/decade, although all are smaller (i.e., less negative). The latter result is due (at least in part) to the larger Pinatubo signal in GISS than in NCEP (Figure 2d). These differences in the GISS and NCEP p LRT responses to Pinatubo eruption may be driven by pro-

6 ACL 1-6 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT Figure 2. Time series of global-mean monthly mean changes in p LRT, the pressure of the lapse rate tropopause, in reanalyses (panel a) and model experiments performed with the ECHAM and GISS models (panels b and c, respectively). A decrease in p LRT signifies an increase in the height of the tropopause. For sources of data, refer to section 2. The top three panels show both low-pass filtered and unfiltered results (bold and thin lines, respectively). The bottom panel (panel d) compares low-pass filtered results in models and reanalyses. Data are expressed as anomalies relative to climatological monthly means over 1979 to 1997 (1979 to 1993 for ERA). Low-pass filtering uses the SMOOFT routine of Press et al. [1986]. nounced differences in both the vertical and the meridional structure of the atmospheric temperature changes after Pinatubo. [38] Tropical p LRT data also show an overall decrease in tropopause pressure in both reanalyses and the ECHAM experiments (Figure 3). The decreases are consistently smaller in the tropics than in the globally averaged data: tropical p LRT trends in GSO1, GSO2, GSOP, and NCEP over are 1.65, 1.50, 0.34, and 1.05 hpa/ decade, respectively. Possible reasons for the smaller tropical trends are discussed in section 5. As in the case of the global-mean data, GSO1 and GSO2 have larger negative trends than NCEP due to their neglect of volcanic effects on p LRT. GSOP has a much smaller negative trend than NCEP due to the unrealistically large tropical response to Pinatubo in the model experiment and the neglect of El Chichón. While NCEP s response is roughly 20% smaller in the tropics than in the global average, the converse is the case for GSOP: the tropical response is about 70% larger than the global-mean response. [39] Several previous studies provide useful checks on our estimates of p LRT changes in the NCEP and ERA reanalyses. Seidel et al. [2001], using three different networks of radiosonde stations located between 15 N 15 S, found least-squares linear trends in p LRT of roughly 0.6 to 0.8 hpa/decade over the period To facilitate comparison with Seidel et al., we averaged NCEP p LRT data over the same latitude band. Our estimated trend over , 0.68 hpa/decade, is consistent with Seidel et al. s results.

7 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT ACL 1-7 := Figure 3. As for Figure 2, but for p LRT values averaged over 30 N 30 S. [40] In contrast, Randel et al. [2000] obtain a smaller trend ( 0.32 hpa/decade over ) using NCEPbased estimates of p LRT averaged over 20 N 20 S. Our estimate of the NCEP trend over the same area and time period is larger than this ( 0.74 hpa/decade) and is in better accord with the Seidel et al. [2001] results. [41] A further reference point is the study by Hoinka [1998], who calculated p LRT from ERA. Hoinka s estimate of the global-mean trend in p LRT ( 0.1 hpa/decade over ) is an order of magnitude smaller than our own estimate from the same data ( 1.13 hpa/decade). This discrepancy may be related to the differences in the algorithms used to determine p LRT (see section 2.3). We note that our result for ERA is in very good agreement with our estimate of the global-mean p LRT trend in the NCEP data ( 1.05 hpa/decade over ) Lower Stratospheric Temperature [42] Lower stratospheric temperatures in actual MSU data and reanalyses are characterized by pronounced decreases in temperature over the satellite era (Figure 4a). Some portion of this multidecadal cooling is attributable to the combined effects of stratospheric ozone depletion and increases in well-mixed greenhouse gases [Ramaswamy et al., 1996, 2001]. Superimposed on the overall cooling of the lower stratosphere are the prominent warming signatures of El Chichón and Pinatubo. The warming results from increased absorption of incoming solar radiation and upwelling terrestrial infrared radiation by volcanic aerosols [Ramaswamy et al., 2001]. [43] Note that the linear decrease in stratospheric temperature over is larger in the NCEP reanalysis than in the actual MSU data ( 0.68 versus 0.53 C/decade, respectively). The cooling of NCEP s equivalent channel 4 temperature relative to MSU begins around the time of the Pinatubo eruption, and is sustained through the end of 2000 (Figure 4a). This discrepancy is partly related to NCEP s assimilation of satellite-based temperature retrievals from the National Environmental Satellite Data and Information Service (NESDIS). These retrievals use information from three different satellite-borne instruments, including MSU. Biases and temporal inhomogeneities in the NESDIS retrievals [Mo et al., 1995; Pawson and Fiorino, 1998; Santer et al., 1999] and changes in the retrieval algorithm [Basist and Chelliah, 1997] contribute to the cooling of NCEP s lower stratosphere relative to MSU. Biases in the radiosonde data assimilated by NCEP may also explain some of the channel 4 trend differences in NCEP and MSU [Lanzante et al., 2002]. [44] The simulated lower stratospheric warming response to Pinatubo in GSOP is roughly twice as large as observed (Figures 4b and 4c). This is partly due to two factors: the model s neglect of the enhanced stratospheric ozone deple-

8 ACL 1-8 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT Figure 4. Time series of global-mean monthly mean changes in stratospheric temperatures in observations/reanalyses (panel a) and in ECHAM model experiments (panel b). Sources of data are given in section 2. Results are for actual MSU channel 4 temperatures, and for equivalent channel 4 temperatures computed from the Bengtsson et al. [1999] GSO experiments and the NCEP and ERA reanalyses. Panel c compares low-pass filtered results in observations, reanalyses, and model experiments. Refer to Figure 1 for information on filtering and definition of anomalies. tion caused by Pinatubo, and failure to simulate the equatorial cooling induced by an easterly phase of the quasibiennial oscillation (QBO) [Solomon et al., 1996; Bengtsson et al., 1999; Kirchner et al., 1999] Tropospheric Temperature [45] Mid- to upper tropospheric temperatures (Figure 5) and lower tropospheric temperature (Figure 6) show pronounced warming in the model experiments and virtually no change or a net cooling in the MSU observations and reanalyses. The differences between lower tropospheric temperature trends in MSU and the three GSO experiments are in part related to the effects of volcanic eruptions and ENSO variability. Since the ECHAM experiments were performed with a coupled model (rather than with prescribed SSTs), the phasing of ENSO-induced variability is different in the model and observations. The neglect of volcanic effects (in GSO1 and GSO2) and of El Chichón (in GSOP) also hampers direct model-versus-observed trend comparisons. [46] Santer et al. [2001] and Wigley and Santer (submitted manuscript, 2002) found that statistical removal of volcano and ENSO effects consistently reduced the differences between observed 2LT trends and those simulated by ECHAM, but did not completely reconcile these differences. Remaining discrepancies are probably due to a combination of model error, inaccurate or missing forcings in the model simulations, and observational error [Wentz et al., 2001]. [47] The cooling of NCEP s lower stratospheric temperatures relative to MSU (by 0.15 C/decade; see section 4.1.2) is also evident in channel 2, and to a lesser extent in the 2LT retrieval. The NCEP channel 2 trend over is negative, while the corresponding trend in the actual MSU data is positive ( and C/decade, respectively). This discrepancy may be partly related to the previously described biases in NCEP s lower stratospheric temperatures. These propagate into the upper troposphere, since channel 2 has a small (roughly 10%) contribution from the lower stratosphere Statistical Significance of Trend Differences [48] We assess the statistical significance of differences between simulated and observed trends as described by Santer et al. [2001]. Our assessment is hampered by the fact

9 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT ACL 1-9 Figure 5. As for Figure 4, but for time series of global-mean monthly mean changes in mid- to upper tropospheric temperatures (MSU channel 2). that the ECHAM4/OPYC experiments are single realizations (with the exception of GSO1 and GSO2), thus yielding relatively noisy estimates of the underlying responses to the imposed forcings. This problem is exacerbated by the relatively short analysis period: all trend comparisons involve global-mean monthly mean anomaly data over January 1979 to December 1997, the period common to observations and model experiments. (ERA results are excluded since the reanalysis ends in February 1994, and GISS trends will be analyzed elsewhere.) Unforced variability can contribute substantially to short-timescale trends, making it difficult to assess the true trend correspondence between models and observations [Santer et al., 2001]. We compare only raw trends here, since our focus is on exploring the effects of different forcings on model/data trend agreement rather than on a comprehensive investigation of trend uncertainties arising from uncertainties in the statistical removal of unforced variability (section 4.1.3). [49] Except for the GSD experiment, none of the simulated p LRT trends differs significantly from the NCEP p LRT trend (at the 5% level or better; Figure 7a). The model integrations show some correspondence between the magnitude of trends in tropospheric temperature and in p LRT. Larger positive trends in channel 2 and 2LT generally correspond to larger decreases in p LRT (i.e., to larger increases in tropopause altitude). This is why GSD, with the smallest increase in lower tropospheric temperature over , has the smallest increase in tropopause height (compare Figures 7a and 7d). [50] The model experiments of Thuburn and Craig [1997] indicate that p LRT changes are related to variations in stratospheric ozone, presumably through ozone-induced stratospheric temperature changes. Experiments that include stratospheric ozone depletion (GSOP, GSO1, and GSO2) have larger cooling of the lower stratosphere than experiments that do not include this effect (GHG, GSD, and GSDIO; Figure 7b). In general, larger cooling of channel 4 temperatures also leads to a larger decrease in p LRT. [51] Stratospheric temperature effects explain why GSO1 and GSO2 have the largest decreases in p LRT. Although their tropospheric temperature trends are smaller than in GHG and GSDIO, their stratospheric cooling is considerably larger. p LRT is thus a measure of the integrated response to temperature changes in the troposphere and stratosphere. GSOP, although it includes stratospheric ozone depletion, additionally incorporates the effects of Pinatubo, which increases p LRT (Figure 2b). It therefore has a smaller negative p LRT trend than either GSO1 or GSO2. [52] Channel 4 cooling trends in all six ECHAM experiments are significantly smaller than in MSU (Figure 7b). As

10 ACL 1-10 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT Figure 6. As for Figure 4, but for time series of global-mean monthly mean changes in lower tropospheric temperatures (MSU 2LT retrieval). noted previously, the GSO1 and GSO2 comparisons have a bias toward increased cooling of the lower stratosphere due to their neglect of volcanically induced stratospheric warming (Figure 4). This is partly offset by a bias toward reduced cooling resulting from the linearized ozone forcing applied by Bengtsson et al. [1999]. [53] There are significant differences between the raw observed and simulated trends in tropospheric temperature (see Santer et al. [2001] and section 4.1.3). This holds for all six experiments, and for both channel 2 and the 2LT retrieval. GSOP, with the most realistic mixture of forcings, yields tropospheric trends closest to those of the actual MSU data. [54] One puzzling feature of Figure 7 is the relatively close correspondence between the p LRT trends in NCEP and the ECHAM simulations. How can this close agreement occur if tropospheric temperatures over are cooling in NCEP, yet warming in all model experiments? The agreement may be partly related to stratospheric cooling, which is significantly larger in NCEP than in the actual MSU data or the ECHAM simulations (Figure 7b). This cooling, like the warming of the troposphere, tends to reduce p LRT, and hence compensates for the different behavior of tropospheric temperatures in NCEP and ECHAM. The implication is that the similar global-mean p LRT trends in NCEP and ECHAM may have different relative contributions from temperature changes in the stratosphere and troposphere. We explore this issue further in section Simulated Temperature and Tropopause Height Changes on Century Timescales [55] To gain a better understanding of the factors that influence p LRT, it is useful to examine the evolution of global-mean p LRT and equivalent MSU temperatures in ECHAM simulations of 21st century climate (Figure 8). In the GHG, GSD, and GSDIO experiments, there is an inverse relationship between changes in tropospheric temperature and p LRT : larger increases in temperature lead to larger decreases in p LRT and increases in tropopause height. After roughly 1990, the signals in p LRT and tropospheric temperatures evolve almost linearly and are clearly separated from the natural internal variability of the ECHAM and PCM control runs. [56] The behavior of stratospheric temperature is more complex (Figure 8b). Equivalent channel 4 temperatures in GHG, GSD, and GSDIO initially decrease from 1860 to present. This is in accord with other model calculations of the expected cooling due to changes in atmospheric CO 2. [Ramaswamy et al., 1996, 2001]. As discussed by Ramaswamy et al. [2001], this CO 2 -induced cooling is actually the result of two competing effects: increases in both the

11 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT ACL 1-11 Figure 7. Simulated and observed trends in p LRT (panel a), and in MSU temperatures for channel 4 and 2 and the 2LT retrieval (panels b, c, and d, respectively). Least-squares linear trends are computed over using globalmean monthly mean anomaly data, with anomalies defined relative to climatological monthly means. The ±2 s confidence intervals are also shown. The grey shading denotes the 2 s confidence intervals for the NCEP p LRT trend and the actual MSU trends. All confidence intervals are adjusted for temporal autocorrelation effects as in Santer et al. [2001], with the exception of lower stratospheric temperature trends. The latter have very high lag-1 autocorrelations and small (5) effective sample sizes. The statistical significance of differences between modeled trends and observations/reanalysis is assessed as described in section 4.2. thermal infrared emissivity and the thermal infrared absorptivity of the lower stratosphere (i.e., approximately hpa; the region spanning the bulk of the channel 4 weights; see Figure 9). The lower stratosphere is cooled by the former effect and warmed by the latter, and the increased emissivity dominates initially. [57] After circa 2030, equivalent channel 4 temperatures begin to warm in the GHG, GSD, and GSDIO integrations. There are several explanations for this behavior. The key issue is the shape of the channel 4 weighting function relative to the profile of the emissivity/absorptivity balance, which affects the distribution of cooling and heating in the stratosphere. In the absence of any other factors, trace gas concentration changes would affect stratospheric temperatures through fast radiative adjustment, which takes place on timescales of 100 days or less. In reality, this rapid radiative adjustment is convolved with the stratosphere s much slower response to multi-decadal anthropogenic changes in the temperature of the surface and troposphere. This slow CO 2 -induced warming of the troposphere leads to increased emission of upwelling longwave radiation, which tends to reduce the initial fast radiative cooling of the lower stratosphere [Forster et al., 1997; Ramaswamy et al., 2001]. [58] A second explanation is that the ECHAM climate change experiments explicitly included changes in CH 4, N 2 O, CFCs, and HCFCs [Roeckner et al., 1999; Bengtsson et al., 1999]. The radiative effects and horizontal and vertical distributions of these trace gases differ from those of CO 2, both initially and in terms of their changes over time. Calculations with a radiative-convective model show that the lower stratospheric cooling resulting from increases in CO 2 is strongly reduced by additionally including changes in CH 4, N 2 O, and CFCs [Ramaswamy et al., 1996]. Experiments with full AGCMs yield similar findings [Wang et al., 1991; Govindasamy et al., 2001]. Thus the explicit radiative effects of well-mixed, non-co 2 greenhouse gases yield less cooling of the lower (and upper) stratosphere than if these trace constituents are treated with an equivalent CO 2 approximation. [59] A third but minor factor is our neglect of ECHAM data above 30 hpa (data for the 10 hpa level were not archived). Since stratospheric cooling tends to increase with increasing height in the stratosphere [Ramaswamy et al., 2001], our channel 4 temperature-change estimates neglect a small region of substantial cooling. However, the effect of this omission is almost certainly small due to mass weighting. [60] In summary, our results suggest that the lower stratospheric temperature-change signal in response to increases in well-mixed greenhouse gases is unlikely to evolve linearly, or show monotonic cooling. The sign of the projected global-mean change in equivalent channel 4 temperatures in the second half of the 21st century may be highly dependent on the specific mix of trace gases employed in the forcing scenario. These signal uncertainties are compounded by uncertainties in the future evolution of stratospheric ozone, which has a large influence on lower stratospheric temperatures [Ramaswamy et al., 1996]. Scenarios of future stratospheric ozone change were not incorporated in any of the ECHAM experiments Tropopause Height Changes in Relation to Vertical Resolution [61] It is instructive to consider whether the decadal-scale decreases in p LRT in Figure 2 are large relative to the vertical resolution of the AGCMs used to produce these estimates. In the 300-year ECHAM control run, the average pressure of the global-mean annual-mean lapse rate tropopause is at roughly 156 hpa. This was estimated from temperatures archived at discrete pressure levels of 30, 50, 70, 100, 150, 200 hpa, etc. (section 2.2). Thus the vertical resolution of the ECHAM4 AGCM in the vicinity of the global-mean tropopause is roughly 50 hpa (i.e., hpa and hpa). [62] Global-mean p LRT values in the ECHAM GHG experiment decrease by roughly 22 hpa over 1860 to 2100 (Figure 8). This overall change is fairly substantial (44%) when measured against the 50 hpa vertical resolution of ECHAM4 in the region of the tropopause. For a standard

12 ACL 1-12 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT Figure 8. Century-timescale changes in global-mean monthly mean p LRT and layer-average atmospheric temperature. Results are from three climate change experiments and from control runs performed with ECHAM4/OPYC [Bengtsson et al., 1999] and PCM [Washington et al., 2000]. Changes are expressed relative to the smoothed initial state of the experiment (climatological monthly means over ). All anomalies are low-pass filtered. atmosphere, a 22 hpa decrease in the pressure of the tropopause would translate to a tropopause height increase of approximately 950 meters (U.S. Standard Atmosphere, 1976). [63] In NCEP, global-mean p LRT decreased by 2.16 hpa/ decade over 1979 to 2000, corresponding to a linear change in p LRT of 4.8 hpa and a height increase of approximately 200 meters over this 22-year period (section 4.1.1). Like ECHAM4, the assimilation model used by NCEP has vertical resolution of 50 hpa in the region of the globalmean tropopause (section 2.1), so that the overall p LRT change is approximately 10% of the vertical grid spacing at the height of the global-mean tropopause. GISS p LRT changes over are also roughly 10% of the vertical grid spacing (40 to 50 hpa) at tropopause height. [64] In all cases considered here, therefore, the decadaland century-scale global-mean p LRT changes are smaller than the vertical resolution of the model grids at tropopause height. This does not imply that these changes are simply sub-grid-scale noise. The thermal signal that drives the overall p LRT decreases is global in scale, and results in global-scale changes to the static stability of the atmosphere (section 5). Our thermally based algorithm for estimating p LRT is capable of resolving these changes. [65] The accuracy of our estimated p LRT decreases should be explored further in models with significantly higher vertical resolution than those examined here. We note that the vertical resolution of ECHAM and GISS at the height of the tropopause is similar in the tropics but not in the extratropics (section 2.2). The fact that the overall tropopause height changes in ECHAM and GISS are in close agreement, despite differences in vertical resolution, suggests that our estimates of p LRT changes may be relatively insensitive to any resolution-dependent biases in the estimated time-mean p LRT values. 5. Vertical Profiles of Temperature Change [66] It is useful to examine the vertical temperature profiles in ECHAM and NCEP in order to understand how p LRT may change in response to external forcing. We first consider annual-mean vertical temperature profiles in

13 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT ACL 1-13 Figure 9. Global-mean monthly mean temperature changes as a function of height in the ECHAM GHG experiment. Changes are defined as in Figure 8. The global-mean weighting function for MSU channel 4 is shown on the right. The vertical scale for the pressure weights is identical to that used for the GHG temperature changes. The black line denotes the estimated global-mean value of p LRT in GHG. See section 4.3 and Figure 12 for further details. NCEP and GSOP, the ECHAM experiment with forcings most relevant for comparison with observations (Figure 10). [67] Simulated and observed temperature profiles are very similar at low latitudes. The model cold bias is evident at mid- and high latitudes (above approximately 300 and 400 hpa, respectively [Gates et al., 1999]). The cold bias has no significant effect on the estimated p LRT, except at high latitudes in the Northern Hemisphere. [68] At low latitudes, the location of the tropopause is welldefined in the time-averaged temperature profiles, and estimated values of p LRT are very close to the transition between large negative and large positive lapse rates (Figure 10). Temperature profiles in mid- to high latitudes have much smaller lapse rates in the vicinity of the tropopause. As a result, p LRT estimates poleward of about 30 are more sensitive to the number of layers used in the interpolation of lapse rates, and the (interpolated) level at which the WMO lapse rate criterion is first satisfied does not necessarily correspond to the level at which lapse rate changes sign (section 2.3). [69] We next use the ECHAM GHG experiment to consider the effects of well-mixed greenhouse gases on p LRT. This integration is sufficiently long (240 years) to provide a reasonable estimate of the signal in p LRT (Figure 8a). [70] In general, p LRT changes depend on two factors, both of which have a strong latitudinal dependence: (1) the baseline temperature profile T(z) in the control run, particularly at the pressure levels immediately above and below the tropopause; (2) the magnitude and shape of T(z), i.e., the change in the T(z) profile in response to an external forcing. [71] To understand how these factors operate, consider an idealized case in which T(z) increases by 5 C at all model levels below the baseline tropopause, is zero above the tropopause, and is latitudinally invariant. The geometry of the control run vertical temperature profiles in Figure 11 dictates that adding these idealized T(z) values to T(z) would result in a decrease in p LRT, with larger decreases (i.e., a higher tropopause altitude) in the extratropics than in the tropics. This is what actually occurs in both hemispheres in the GHG experiment (Figure 11). [72] Simulated T(z) profiles in GHG deviate from this ideal case, and there is large latitudinal variability in the relative temperature changes immediately above and below the tropopause (Figure 12). Three aspects of the T(z) results are important in interpreting the p LRT changes: (1) Temperature increases immediately below the diagnosed tropopause are largest in the tropics and smallest in the extratropics. Taken alone, this would tend to yield larger p LRT decreases in the tropics than in the extratropics. (2) The warming immediately above the diagnosed tropopause level is also largest in the tropics. This tends to counteract (1): i.e., it would tend to produce smaller p LRT decreases in the tropics than in the extratropics. (3) The net cooling of the stratosphere is larger in the extratropics than in the tropics. This acts in the same direction as item 2 in terms of its effect on p LRT. [73] These complexities in T(z) are partly due to the fact that we are dealing with changes in a number of different greenhouse gases. Each of these has different spatio-temporal distributions, radiative properties, and atmospheric temperature responses (see Govindasamy et al. [2001] and section 4.3). These differences in T(z), coupled with latitudinal variations in T(z), make it difficult to determine a priori how a combination of forcings from CO 2 and non- CO 2 greenhouse gases will affect p LRT. [74] Nevertheless, items 2 and 3 act in the same sense as the latitudinal differences in T(z). All tend to generate larger p LRT

14 ACL 1-14 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT found close to the locations of the subtropical jets (at roughly 30 N and 30 S), close to the boundary between the tropical and extratropical tropopause. Since p LRT gradients are very steep at this boundary (Figure 1), relatively small changes in the position or intensity of the jets can lead to large changes in p LRT. [77] GSDIO, which lacks stratospheric ozone depletion, has smaller p LRT decreases in the extratropics than the three experiments that include this effect. This suggests that the p LRT decrease caused by the larger stratospheric cooling in the GSO integrations outweighs the p LRT decrease caused by the larger tropospheric warming in GSDIO (see Figure 7 and section 4.2). [78] When the large spatial mean terms are subtracted from the patterns of annual-mean p LRT change, the NCEP and ECHAM spatial anomaly fields are virtually uncorrelated ( 0.09 to 0.08; Table 1). There is little indication that increasing realism in the forcing yields greater similarity between the ECHAM and NCEP p LRT changes. We explore this issue in more detail in the paper by Santer et al. (submitted manuscript, 2002) Stratospheric Temperature [79] The simulated changes in stratospheric temperature in the three GSO experiments show strong meridional gra- Figure 10. Profiles of annual-mean temperature as a function of height in the NCEP reanalysis and the ECHAM GSOP experiment (section 2). Results are for high-, midand low latitudes of the Northern and Southern Hemispheres (panels a and b) and are climatological averages over The centers of pressure levels are marked. The solid (dashed) horizontal lines give the location of estimated p LRT values in the GSOP (NCEP) data. decreases in the extratropics than in the tropics. Their net influence apparently outweighs the effect of item 1 above. [75] We make no attempt here to isolate the effects of volcanoes and stratospheric ozone depletion on vertical temperature profiles, and hence on p LRT. The principal constraint is that the GSO integrations involve simultaneous changes in a number of different forcings (section 2.2). We note, however, that as in the GHG case, ozone-induced cooling of the stratosphere is largest at high latitudes (section 6.2). Stratospheric ozone depletion therefore acts in the same direction as the baseline geometry effect, and reinforces the tendency of well-mixed greenhouse gases to cause larger increases in tropopause height in the extratropics. 6. Horizontal Patterns 6.1. Lapse Rate Tropopause [76] Patterns of annual-mean p LRT changes over 1979 to 1997 are similar in the NCEP reanalysis and the ECHAM climate change experiments (Figure 13). Both have areas of p LRT decrease that are much larger than the areas of p LRT increase. The largest decreases in p LRT are at mid- to high latitudes in both hemispheres, particularly in NCEP and the GSO experiments (see section 5). Areas of increase are Figure 11. Profiles of annual-mean temperature as a function of height in the ECHAM control run and GHG experiment. Results are climatological averages over (control) and (GHG). The change in p LRT between the average control state and the final 50 years of GHG is indicated on the right-hand side of each panel. See Figure 10 and section 5 for further details.

15 SANTER ET AL.: BEHAVIOR OF TROPOPAUSE HEIGHT ACL N 45 S (Figures 14e and 14f ). Possible explanations for these differences were discussed in section [82] The pattern correlation results confirm that stratospheric temperature changes in GSDIO are dissimilar to the changes in either NCEP or MSUd (Table 1). Spatial anomaly correlations are consistently negative for NCEP/ GSO comparisons and positive for MSUd/GSO comparisons. Most of this sensitivity to the choice of observed data set probably arises from the excessive cooling of NCEP s stratospheric temperatures in the tropics. In comparisons between GSO and MSUd, greater realism in the forcing enhances the similarity between simulated and observed temperature-change patterns. Figure 12. Profiles of changes in annual-mean temperature as a function of height in the ECHAM GHG experiment. Changes are defined as the differences between the smoothed final and initial states of GHG ( minus ). Climatological-mean p LRT values for and are indicated by the lower and upper limits (respectively) of the shaded horizontal bars. dients, with maximum cooling in polar regions (Figure 14). This is where the largest decreases in stratospheric ozone have occurred over [Hansen et al., 2002]. Over the Arctic, GSO1 cools by up to 4 5 C more than GSO2, highlighting signal uncertainties arising from internally generated atmospheric variability. GSOP s pattern of change in channel 4 temperature is similar to those of GSO1 and GSO2, since the five-year means that are differenced ( minus ) largely exclude Pinatubo effects. The cooling resulting from stratospheric ozone depletion is readily apparent by comparing GSDIO results with those of the GSO integrations. [80] MSU channel 4 data also show the largest cooling in polar regions, but have bands of subsidiary cooling maxima centered at 30 N and 30 S. These bands are absent in the GSO integrations, but do appear in the NCEP data. The global-mean cooling is considerably larger in NCEP and MSUd than in the GSO integrations (Figure 7b). Unlike the p LRT results, increasing realism of the forcing does translate to greater similarity between modeled and observed globalmean changes. [81] The NCEP and MSU patterns of stratospheric temperature change are very similar poleward of 60, but NCEP s cooling substantially exceeds that of MSU between roughly 6.3. Midtropospheric to Upper Tropospheric Temperature [83] Simulated temperature changes in the mid- to upper troposphere (Figure 15) also show a clear signal in response to the imposed stratospheric ozone depletion in the GSO integrations. Equivalent channel 2 temperatures in GSO1, GSO2, and GSOP cool by up to 0.6 to 0.8 C over the Arctic and Antarctic. This polar cooling is absent in the GSDIO integration without ozone depletion, but is present in actual MSU channel 2 data. [84] The most prominent differences between mid- to upper-tropospheric temperatures in the GSO experiments and MSU are between ca. 30 N 30 S, where the simulations warm and MSU cools. NCEP s cooling over the same region is 0.2 to 0.4 C larger than in MSU. This is reflected in the large differences in global-mean channel 2 trends in NCEP and MSU (see Figure 7c and section 4.1.3). In common with results for stratospheric temperatures, GSO spatial anomalies are more similar to MSUd than to NCEP (Table 1 and section 6.2). A further commonality is that the pattern of mid- to upper-tropospheric temperature change in MSUd is more highly correlated with the GSO anomaly fields (with more realistic forcing) than with the GSDIO pattern Lower Tropospheric Temperature [85] Lower tropospheric temperature changes (Figure 16) are spatially noisier than the temperature changes in channels 2 and 4. As for channel 2, MSU and NCEP 2LT data cool in the tropics and subtropics, while all four model simulations warm. The simulated and observed patterns of 2LT change are virtually uncorrelated (Table 1). The MSU and NCEP temperature-change patterns are more similar for the 2LT retrieval than for channels 4 and 2. [86] There is a tendency toward larger lower tropospheric temperature increases at the poles and a net decrease in the equator-to-pole temperature gradient. Exceptions are GSOP in the Northern Hemisphere and MSUd in the Southern Hemisphere. Held [1982] has suggested that an increase in equator-to-pole temperature gradient might increase baroclinicity and raise the height of the extratropical tropopause. In the model experiments of Thuburn and Craig [1997], the opposite result occurred: increasing (decreasing) the equator-to-pole temperature gradient lowered (raised) tropopause height in the extratropics. [87] The findings of Thuburn and Craig [1997] help to explain two aspects of our results. The first aspect is the previously noted tendency toward larger changes in p LRT at